Beyond the Premier: Assessing Action Spotting Transfer Capability Across Diverse Domains

Abstract

Football stands as one of the most successful sports in history thanks to the plethora of professional leagues broadcasted worldwide followed by avid fans, further fueled by the abundance of amateur and grassroots leagues across nearly every country, encompassing countless players who devote their time to the sport. Despite the tremendous amount of visual data available worldwide for developing automatic systems to extract game events, most efforts focus on the few professional league matches. However, the recording quality and broadcasts editing vary considerably across leagues, creating a disparity in the analytical capabilities of deep learning models. This paper delves into an analysis of how action spotting models transfer to diverse domains, analyzing the performance gap between various types of broadcasts. In particular, we investigate the transfer capability of state-of-the-art action spotting models across leagues, from amateur to professional, and broadcast quality, from AI-piloted camera to professional broadcast editing. Our analysis shows that transferring across leagues is challenging, with the most impactful feature being broadcasting editing quality. This analysis paper therefore seeks to spotlight this pressing issue and catalyze future research endeavors in the field of domain adaptation for action spotting methods.